AQI Prediction With Spatial-Temporal Deep Neural Network And Hyper-Parameters’ Optimization

The frequent occurrence of heavy air pollution has caused serious harm to people’s life and health.The timely and accurate prediction of air quality can not only help people understand the future air conditions in advance,but also provide early warning before the heavy pollution weather occurred,so that the environmental protection departments can take emergency response to prevent or control the occurrence of hazards.Air quality prediction has great practical significance,but the prediction accuracy is still need to be improved.Neural networks have been shown to be used in sequential data prediction problems.Up to now,there have been few applications of deep learning in the field of air quality prediction.The work has rarely considered the spatio-temporal characteristics of air quality when selecting models,and it is too enpirical when selecting network hyper-parameters.In order to solve the above problems,we take the network structure,hyper-parameter selection and model input as the three entry points for the research and propose two models for air quality prediction:To solve the problem of low accuracy and sensitive to noisy of existing air quality prediction models,we proposed an air quality prediction model SDAE-GS,which is based on stacked denoising auto-encoders(SDAE)and grid search method.Firstly,the historical air quality and meteorological monitoring data of Wuhan city are taken as research object.SDAE model is established to study the characteristic expression of the original data layer by layer,and the last layer is connected with a BP layer to tune the prediction model.Next,we design a multidimensional grid quadratic search method to perform hyperparameter optimization and make predictions on the test set.The average absolute error,root mean square error,and mean absolute percentage error between the predicted value and the actual value are used as prediction performance evaluation standards.Compared with other network models,it can be proved that SDAE model has better predictive performance.Finally,the input data is optimized considering their spatial and temporal relations.Experimental results show that the spatial optimization based SDAE has the most improvement for predictive performance,and it can obtain more accurate predictions compared with the traditional methods.Due to the SDAE-GS model has no time modeling capability,we proposed another air quality prediction model LSTM-FWA,which is based on long-short term memory network and fireworks algorithm.Firstly,the time variation characteristics of air quality are modeled by right of the structure of interconnections between hidden layer nodes in LSTM network.Then the swarm intelligent algorithm is used to improve the hyper-parameters optimization.Considering the population diversity and concurrency,fireworks algorithm is applied to the hyper-parameters optimization.Finally,the model input is optimized in temporal,spatial and spatio-temporal respectively,experimental results show that both spatio-temporal and temporal optimization based LSTM has the more obvious improvement for predictive performance.When the work finished,we compare SDAE-GS with LATM-FWA and get the comprehensive analysis of the performance about the two models under various optimization strategy,the experimental results show that the LSTM-FWA model under the spatio-temporal optimization strategy has better prediction performance for air quality.